AI FOR SCIENCE · INTERVIEWS
Two Anthropic leads say Claude should be a research assistant across the whole workflow: first connect to the tools scientists use daily, then reach experiment execution, then learn from lab data.
Names the base-layer partners (Benchling, 10x Genomics, PubMed, with Sage Bionetworks and BioRender also mentioned), calls Sonnet 4.5 the first model with extensive scientific training, and puts Claude running lab experiments and learning actively from high-throughput data as the next steps.
The interview gives no figures, partner results or lab-execution demonstration, and its only case is a personal anecdote; the same-day announcement gives a Protocol QA score, and later evidence is under After the interview.
The first layer is connecting to daily tools: experiment records (Benchling), single-cell analysis (10x Genomics' Cell Ranger) and literature (PubMed). The second is a research assistant across hypothesis, experiment execution (drafting and debugging protocols), computational analysis and write-up, with each task type broken out, evaluated and improved.
Kauderer-Abrams says Sonnet 4.5 is Anthropic's first model with extensive scientific training, and that better maths lifts the computational side of biology; more important is long strings of tool calls, which bioinformatics pipelines need. He adds that more and more people in the community use Claude Code for bioinformatics.
They say the target is the whole span from early discovery to development and translation. Drawing on his own experience writing regulatory submissions and dealing with the FDA, Cool says Claude is capable there and could speed up both industry and the FDA.
He says a growing number of papers suggest tasks that once needed specialised biology models may be done by large frontier models with the right training, and that Anthropic will pursue this aggressively. He also grants that those models' superhuman abilities need a language interface to be usable, and that the field is at the beginning.
Kauderer-Abrams says Claude must learn to execute experiments in the lab: design a plan, draft protocols, then run them while he reviews data in the morning; and to do lab-in-the-loop active learning, because learning from human experts will saturate while experimental throughput keeps scaling each year.
“At some point we're going to saturate learning from human experts. The answer is to get the data from the lab.”Eric Kauderer-AbramsAnthropic's announcement of the same day (2025-10-20) already says the goal is to support the whole process from early discovery to translation and commercialisation; Sonnet 4.5 scores 0.83 on Protocol QA (human baseline 0.79, Sonnet 4 at 0.74); new connectors include Benchling, PubMed and 10x Genomics, with Agent Skills; and it says eventually AI models should make such discoveries autonomously.
The two leads put the route in order: first connect to the tools scientists use daily, then a research assistant across hypothesis, protocol and analysis, then Claude executing experiments in the lab, and last learning actively from high-throughput lab data.
The announcement describes the software layer: connectors, skills, a prompt library, dedicated support and consulting partners. It does not describe execution in a lab or active learning.
The interview names Claude running lab experiments and learning actively from lab data as the next steps, and gives the north star as speeding up life-sciences R&D by at least an order of magnitude.
The product is connectors plus long tool-call chains, not a new science model, with the difficulty placed in tool integration and task evaluation. Whoever holds the lab record and data entry point holds the door on this route; that is our inference.
Kauderer-Abrams says a growing number of papers suggest large models with the right training may replace some specialised biology models, and Anthropic will push on it. He names no papers, so it is a direction, not a tested comparison.
Sonnet 4.5 is Anthropic's first model with extensive scientific training, with a big jump in long tool-call tasks.
Company-reported
"First with extensive scientific training" is Kauderer-Abrams' statement and the interview gives no figures; the same-day announcement gives Protocol QA at 0.83 (human baseline 0.79).
Connected to Benchling, 10x Genomics and PubMed, with Sage Bionetworks and BioRender also early partners.
Company-reported
The same-day announcement confirms connectors for Benchling, PubMed and 10x Genomics, with Sage's Synapse.org and BioRender also listed.
Kauderer-Abrams' biotech's COVID assay was inhibited by the sample matrix and took three months to fix; Claude gave an answer in one minute, naming what to add.
Company-reported
A personal anecdote; he says the answer was imperfect at the time, and whether it was right was not checked.
Claude executing experiments in the lab, and lab-in-the-loop active learning.
Target
A direction, with no date, partner or result.
The north star is accelerating life-sciences R&D by at least an order of magnitude.
Target
Taken from Dario Amodei's essay; no way of measuring it is given.
Papers suggest general frontier models may replace some specialised biology models.
Industry context
The guest's reading; no papers are named.
The interview gives no benchmark figures, partner results or lab-execution demonstration.
Not yet closed
Partnerships such as the Arc Institute are named without detail.
Whether the tool layer becomes one working environment scientists use end to end.
Would show up as A released product that integrates the tools and keeps an auditable record, plus independent use reports.Partly answered Claude Science was released in beta on 2026-06-30, integrating common tools and keeping auditable artifacts (Anthropic's announcement). The first reports from outside users come from a week-long July hackathon Anthropic ran with the Gladstone Institutes: winners used Claude to rank variants, design enzymes and revive old code, but the hosts judged, none is peer-reviewed or lab-tested, and 5 of 20 sampled revived repositories reproduced their paper's numbers. On 8 October two preprints by outside researchers each reported a workflow Claude ran end to end: Academia Sinica researchers had Claude Code supervise model building and refinement for 24 structures, steering throughout, and IFP Energies nouvelles and Imperial College researchers had Claude derive, code and fit an equation of state, with tests written by the same model. Both are the authors' own reports, neither mentions Claude Science, and neither compares with human experts or other tools.Whether the per-task evaluations are published, with figures.
Would show up as An Anthropic life-sciences benchmark or system card.Partly answered The same-day announcement published a Protocol QA score (0.83, human baseline 0.79). BioMysteryBench, of 2026-04-29, is an Anthropic evaluation with figures for one of the task types, computational analysis on real data: 99 questions, and a 30% solve rate for Claude Mythos Preview on the 23 no expert on the panel solved (on the timeline since 24 October). Anthropic built it and scored only its own models, and for the two other task types the interview lists, hypotheses and write-up, no published evaluation has been seen.Whether experiments Claude runs in the lab produce results beyond the first report, checked outside.
Would show up as A second and later wet-lab result stating which steps humans did.Partly answered The first result, of 2026-09-23, reached transcription only, and the bioRxiv preprint of 11 October, with Kauderer-Abrams among its authors, writes it up with no new wet-lab work: ten reruns of the same campaign never found the array again. The scorpion-toxin binder campaign published on 20 October is the second result with wet-lab data, and the measuring was done outside: two contract research organisations tested 211 designs and 17 met the company's criteria (on the timeline since 22 October). So the outside-measurement half now exists; the autonomy half is still the company's own account, with the criteria set after the data came in, no baseline run, and the company itself saying it has produced no drug. On 27 October Anthropic's post reported three more wet-lab results, from partners' labs, each saying which steps people did: six Generate scientists used Claude Code to propose CAR signalling-tail designs, tested in Generate's own lab; BigHat's lab made and tested about 1,010 variants of the changes Claude proposed, with BigHat's scientists refining and running the revised protocol; and at Divergence Labs Claude generated about 35,000 antibody designs largely on its own, which the company narrowed down and tested (on the timeline since 27 October). Every measurement was made in the partner's own lab and every report is the company's own, not peer reviewed, so the outside check still goes no further than the partner. Divergence's technical report (on the timeline since 29 October) gives the half Anthropic's post left out: of 91 designs, 20 beat every negative control, and of the eight titrated, three removed more IgE than the approved drug in the same assay, but the report also says the variants did not appear to improve on the parent molecule they started from, and every measurement was on a sensor chip, with no cells or animals.Whether a general model matches a specialised biology model on the same task.
Would show up as An independent paper with the same task and benchmark.Whether the AI for Science programme and partners such as the Arc Institute and Novo Nordisk publish results.
Would show up as Research outputs published by the programme or a partner.Later evidence that tests or limits what was said, and targets whose date has passed.
Anthropic released Claude Science in beta: one app that integrates the tools and packages researchers use, keeps an auditable history of how each result was made, and connects to an HPC cluster or on-demand GPUs. It delivers on connecting the tools first and is the first product shape of an assistant across the workflow; the announcement gives no result figures.Anthropic, Claude Science ↗
The wet lab reported its first result: an agent narrowed 198,290 clusters to 19 reports and flagged an unannotated repeat array, and human-run sequencing shows it is transcribed. One run, transcription only, so it is the first evidence on learning from lab data, and a limited one.Timeline entry →
The bioRxiv preprint of this result, with Kauderer-Abrams among its authors, reports that ten reruns of the same agent campaign nearly all reached the ART loci, but none read the DNA upstream and none found the array; on fixed inputs the strongest models described the array in at least 90% of attempts with the DNA in the prompt, and as few as 32% when it came as files to open. That limits the interview's picture of Claude doing research in the lab: the one find depended on how the input reached the model, and did not repeat.Timeline entry →
Read 2026.10.01 The full YouTube auto-captions; names and titles follow the video description. The video is a year old and is recorded as of then. Open the interview ↗